Unable to export graph as ensemble when using multi-hot categorical column

[mkumari-ed]

I am using inference container 22.05. When materializing feast in cell 6 deploying multi-stage recsys notebook, I am getting exception while materializing user_features. Item features worked just fine. Please download the sample data and notebooks used for this from the link here

Using tensorflow-inference 22.05 container. Attaching the error here

[karlhigley]

What Feast version are you using?

[mkumari-ed]

Hi Karl, I am using Feast==0.19.4. I tried removing the multi-hot categorical feature column(for example: search_term, one of the examples being ['tokkio','merlin','riva documentation','nvbot implementation']) and I was able to materialize it without that. @rnyak has informed that your team hasn't been able to test the multi-hot categorical feature columns in the dataset with Feast and that is the reason why it's happening. I would like to not remove this feature column from my dataset as it improves the model performance to some extent. Could you or someone from the team please work on testing it out?

[rnyak]

Closing this since materialization issue was solved. And another bug was reported here NVIDIA-Merlin/systems#115

[mkumari-ed]

@karlhigley and @rnyak
I tested it out with my correct data today. I am attaching all screenshots for visibility. With the list of strings feature(for ex. ['sipl_camera', 'build qnx', 'yuv', 'dl g']), I am unable to materialize the user features using feast. The error that I get is ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all(). Also, attaching the user features sample and output of train_tt.to_ddf().compute() here along with the snapshot of how the raw data looks like

[mkumari-ed]

So going back to our conversation in the past, I used INT32_LIST dtype for this list type column instead of INT32 and I was able to materialize as well as use queryfeast operator in the successive steps. However it leads to another error now on the step where we create retrieval object using PredictTensorflow object.

[karlhigley]

What does model._saved_model_inputs_spec look like for your retrieval model?

[mkumari-ed]

@karlhigley here is the output for the retrieval model
{'search': (TensorSpec(shape=(727237, 1), dtype=tf.int32, name='search_1'), TensorSpec(shape=(16384, 1), dtype=tf.int32, name='search_2')), 'userid': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='userid'), 'contentid': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='contentid'), 'eventtype': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='eventtype'), 'eventStrength': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='eventStrength'), 'ts': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='ts'), 'spaceid': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='spaceid'), 'spacekey': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='spacekey'), 'spacename': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='spacename'), 'parenttitle': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='parenttitle'), 'parentid': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='parentid'), 'mgrntaccount': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='mgrntaccount'), 'businessarea': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='businessarea'), 'jobposition': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='jobposition'), 'hrrolename': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='hrrolename'), 'location': TensorSpec(shape=(16384, 1), dtype=tf.int32, name='location')}